From: Sebastian Dröge Date: Mon, 14 Dec 2009 19:08:06 +0000 (+0100) Subject: videomixer: Add MMX implementations of the AYUV blending and color filling functions X-Git-Tag: RELEASE-0.10.18~183 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=061ededa360bc6da264a6de702157a4792772b5b;p=platform%2Fupstream%2Fgst-plugins-good.git videomixer: Add MMX implementations of the AYUV blending and color filling functions This provides a 20% speedup for blending and 100% for color filling. The blending can probably be optimized even more. --- diff --git a/gst/videomixer/Makefile.am b/gst/videomixer/Makefile.am index 513c88f..e217832 100644 --- a/gst/videomixer/Makefile.am +++ b/gst/videomixer/Makefile.am @@ -1,8 +1,8 @@ plugin_LTLIBRARIES = libgstvideomixer.la libgstvideomixer_la_SOURCES = videomixer.c blend_ayuv.c blend_bgra.c blend_i420.c blend_rgb.c -libgstvideomixer_la_CFLAGS = $(GST_CFLAGS) $(GST_BASE_CFLAGS) $(GST_CONTROLLER_CFLAGS) $(GST_PLUGINS_BASE_CFLAGS) -libgstvideomixer_la_LIBADD = $(GST_LIBS) $(GST_BASE_LIBS) $(GST_CONTROLLER_LIBS) $(GST_PLUGINS_BASE_LIBS) -lgstvideo-@GST_MAJORMINOR@ +libgstvideomixer_la_CFLAGS = $(GST_CFLAGS) $(GST_BASE_CFLAGS) $(GST_CONTROLLER_CFLAGS) $(GST_PLUGINS_BASE_CFLAGS) $(LIBOIL_CFLAGS) +libgstvideomixer_la_LIBADD = $(GST_LIBS) $(GST_BASE_LIBS) $(GST_CONTROLLER_LIBS) $(GST_PLUGINS_BASE_LIBS) -lgstvideo-@GST_MAJORMINOR@ $(LIBOIL_LIBS) libgstvideomixer_la_LDFLAGS = $(GST_PLUGIN_LDFLAGS) libgstvideomixer_la_LIBTOOLFLAGS = --tag=disable-static diff --git a/gst/videomixer/blend_ayuv.c b/gst/videomixer/blend_ayuv.c index 1741c47..17d63d5 100644 --- a/gst/videomixer/blend_ayuv.c +++ b/gst/videomixer/blend_ayuv.c @@ -1,5 +1,6 @@ /* * Copyright (C) 2004 Wim Taymans + * Copyright (C) 2009 Sebastian Dröge * * This library is free software; you can redistribute it and/or * modify it under the terms of the GNU Library General Public @@ -17,9 +18,18 @@ * Boston, MA 02111-1307, USA. */ +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif #include +#ifdef HAVE_GCC_ASM +#if defined(HAVE_CPU_I386) || defined(HAVE_CPU_X86_64) +#define BUILD_X86_ASM +#endif +#endif + #define BLEND_NORMAL(Y1,U1,V1,Y2,U2,V2,alpha,Y,U,V) \ Y = ((Y1*(255-alpha))+(Y2*alpha))>>8; \ U = ((U1*(255-alpha))+(U2*alpha))>>8; \ @@ -230,6 +240,93 @@ gst_videomixer_blend_ayuv_ayuv (guint8 * src, gint xpos, gint ypos, #undef BLEND_MODE +#ifdef BUILD_X86_ASM +void +gst_videomixer_blend_ayuv_ayuv_mmx (guint8 * src, gint xpos, gint ypos, + gint src_width, gint src_height, gdouble src_alpha, + guint8 * dest, gint dest_width, gint dest_height) +{ + gint b_alpha; + gint i; + gint src_stride, dest_stride; + gint src_add, dest_add; + + src_stride = src_width * 4; + dest_stride = dest_width * 4; + + b_alpha = (gint) (src_alpha * 255); + + /* adjust src pointers for negative sizes */ + if (xpos < 0) { + src += -xpos * 4; + src_width -= -xpos; + xpos = 0; + } + if (ypos < 0) { + src += -ypos * src_stride; + src_height -= -ypos; + ypos = 0; + } + /* adjust width/height if the src is bigger than dest */ + if (xpos + src_width > dest_width) { + src_width = dest_width - xpos; + } + if (ypos + src_height > dest_height) { + src_height = dest_height - ypos; + } + + src_add = src_stride - (4 * src_width); + dest_add = dest_stride - (4 * src_width); + + dest = dest + 4 * xpos + (ypos * dest_stride); + + for (i = 0; i < src_height; i++) { + /* *INDENT-OFF* */ + __asm__ __volatile__ ( + "pxor %%mm7 , %%mm7 \n\t" /* mm7 = 0 */ + "pcmpeqd %%mm6 , %%mm6 \n\t" /* mm6 = 0xffff... */ + "punpcklbw %%mm7 , %%mm6 \n\t" /* mm6 = 0x00ff00ff00ff... */ + "pcmpeqd %%mm5 , %%mm5 \n\t" /* mm5 = 0xffff... */ + "psrlq $56 , %%mm5 \n\t" /* mm5 = 0x0...0ff */ + "xor %%ecx , %%ecx \n\t" /* ecx = 0 */ + "1: \n\t" + "movzxb (%0) , %%eax \n\t" /* eax == source alpha */ + "imul %2 , %%eax \n\t" /* eax = source alpha * alpha */ + "sar $8 , %%eax \n\t" /* eax = (source alpha * alpha) / 256 */ + "movd %%eax , %%mm0 \n\t" /* mm0 = apply alpha */ + "movd (%0) , %%mm2 \n\t" /* mm2 = src */ + "movd (%1) , %%mm1 \n\t" /* mm1 = dest */ + "punpcklwd %%mm0 , %%mm0 \n\t" + "punpckldq %%mm0 , %%mm0 \n\t" /* mm0 == 0a 0a 0a 0a */ + "punpcklbw %%mm7 , %%mm1 \n\t" /* mm1 == dv du dy da */ + "punpcklbw %%mm7 , %%mm2 \n\t" /* mm2 == sv su sy sa */ + "pmullw %%mm0 , %%mm2 \n\t" /* mm2 == a * s */ + "pandn %%mm6 , %%mm0 \n\t" /* mm0 == 255 - a */ + "pmullw %%mm0 , %%mm1 \n\t" /* mm1 == (255 - a) * d */ + "paddusw %%mm2 , %%mm1 \n\t" /* mm1 == s + d */ + "psrlw $8 , %%mm1 \n\t" + "packuswb %%mm7 , %%mm1 \n\t" + "por %%mm5 , %%mm1 \n\t" /* mm1 = 0x.....ff */ + "movd %%mm1 , (%1) \n\t" /* dest = mm1 */ + "add $4 , %1 \n\t" + "add $4 , %0 \n\t" + "add $1 , %%ecx \n\t" + "cmp %%ecx , %3 \n\t" + "jne 1b" + : /* no output */ + :"r" (src), "r" (dest), "r" (b_alpha), "r" (src_width) + :"%eax", "%ecx", "memory" +#ifdef __MMX__ + , "mm0", "mm1", "mm2", "mm5", "mm6", "mm7" +#endif + ); + /* *INDENT-ON* */ + src += src_add; + dest += dest_add; + } + __asm__ __volatile__ ("emms"); +} +#endif /* fill a buffer with a checkerboard pattern */ void @@ -263,3 +360,42 @@ gst_videomixer_fill_ayuv_color (guint8 * dest, gint width, gint height, } } } + +#ifdef BUILD_X86_ASM +void +gst_videomixer_fill_ayuv_color_mmx (guint8 * dest, gint width, gint height, + gint colY, gint colU, gint colV) +{ + guint64 val; + guint nvals = width * height; + + val = (((guint64) 0xff)) | (((guint64) colY) << 8) | + (((guint64) colU) << 16) | (((guint64) colV) << 24); + val = (val << 32) | val; + + /* *INDENT-OFF* */ + __asm__ __volatile__ ( + "cmp $2 , %2 \n\t" + "jb 2f \n\t" + "movq %4 , %%mm0 \n\t" + "1: \n\t" + "movq %%mm0 , (%1) \n\t" + "sub $2 , %0 \n\t" + "add $8 , %1 \n\t" + "cmp $2 , %2 \n\t" + "jae 1b \n\t" + "emms \n\t" + "2: \n\t" + : "=r" (nvals), "=r" (dest) + : "0" (nvals), "1" (dest), "r" (val) + : "memory" +#ifdef __MMX__ + , "mm0" +#endif + ); + + /* *INDENT-ON* */ + if (nvals) + GST_WRITE_UINT32_LE (&dest[-4], (guint32) (val & 0xffffffff)); +} +#endif diff --git a/gst/videomixer/videomixer.c b/gst/videomixer/videomixer.c index 4599760..f404c28 100644 --- a/gst/videomixer/videomixer.c +++ b/gst/videomixer/videomixer.c @@ -57,11 +57,21 @@ #include "config.h" #endif +#ifdef HAVE_GCC_ASM +#if defined(HAVE_CPU_I386) || defined(HAVE_CPU_X86_64) +#define BUILD_X86_ASM +#endif +#endif + #include #include #include #include +#include +#include +#include + #ifdef HAVE_STDLIB_H #include #endif @@ -160,6 +170,14 @@ void gst_videomixer_fill_i420_checker (guint8 * dest, gint width, gint height); void gst_videomixer_fill_i420_color (guint8 * dest, gint width, gint height, gint colY, gint colU, gint colV); +#ifdef BUILD_X86_ASM +void gst_videomixer_blend_ayuv_ayuv_mmx (guint8 * src, gint xpos, gint ypos, + gint src_width, gint src_height, gdouble src_alpha, + guint8 * dest, gint dest_width, gint dest_height); +void gst_videomixer_fill_ayuv_color_mmx (guint8 * dest, gint width, gint height, + gint colY, gint colU, gint colV); +#endif + #define DEFAULT_PAD_ZORDER 0 #define DEFAULT_PAD_XPOS 0 #define DEFAULT_PAD_YPOS 0 @@ -920,12 +938,25 @@ gst_videomixer_setcaps (GstPad * pad, GstCaps * caps) goto done; switch (mixer->fmt) { - case GST_VIDEO_FORMAT_AYUV: + case GST_VIDEO_FORMAT_AYUV:{ +#ifdef BUILD_X86_ASM + guint cpu_flags = oil_cpu_get_flags (); + + mixer->blend = + (cpu_flags & OIL_IMPL_FLAG_MMX) ? gst_videomixer_blend_ayuv_ayuv_mmx : + gst_videomixer_blend_ayuv_ayuv; + mixer->fill_checker = gst_videomixer_fill_ayuv_checker; + mixer->fill_color = + (cpu_flags & OIL_IMPL_FLAG_MMX) ? gst_videomixer_fill_ayuv_color_mmx : + gst_videomixer_fill_ayuv_color; +#else mixer->blend = gst_videomixer_blend_ayuv_ayuv; mixer->fill_checker = gst_videomixer_fill_ayuv_checker; mixer->fill_color = gst_videomixer_fill_ayuv_color; +#endif ret = TRUE; break; + } case GST_VIDEO_FORMAT_I420: mixer->blend = gst_videomixer_blend_i420_i420; mixer->fill_checker = gst_videomixer_fill_i420_checker; @@ -1623,6 +1654,8 @@ plugin_init (GstPlugin * plugin) GST_DEBUG_CATEGORY_INIT (gst_videomixer_debug, "videomixer", 0, "video mixer"); + oil_init (); + return gst_element_register (plugin, "videomixer", GST_RANK_PRIMARY, GST_TYPE_VIDEO_MIXER); }